Add challenge 111: Split-KV Attention Reduction (Medium)

[claude]

Summary

• New medium challenge: a Flash-Decoding-style merge kernel that combines partial attention outputs from multiple KV-cache chunks into the final attention result.
• Teaches the log-sum-exp merge trick used to safely combine independently-normalized softmax outputs — a core primitive in production LLM inference stacks (vLLM, FlashInfer) for long-context decoding.
• Inputs are per-split (partial_out, partial_lse) tuples; the solver computes stable weights, normalizes, and reduces along the split axis.

Validation

• Verified with run_challenge.py --action submit on Tesla T4 — all tests passed.
• pre-commit run --all-files passes clean (black, isort, flake8, clang-format, mojo format).

Test plan

• challenge.py reference matches example output [[2, 3], [3, 2]]
• Functional tests cover single-split, all-equal-lse, zero partial_out, non-power-of-2 splits/heads/dims, and realistic sizes
• Performance test (num_splits=128, num_heads=64, head_dim=128) fits well within Tesla T4 memory
• All 6 starter files present and compile without solving the problem
• Linting passes

🤖 Generated with Claude Code

[kunal-mansukhani]

Thanks for the PR! Challenge ID 111 is already taken on main (challenges/medium/111_*). Please renumber to the next free ID (currently unused near the top of the range: 120+, or any gap such as 95/97/98/99/100/101/102/103/104) and update the directory name + PR title accordingly. Happy to merge once the ID no longer collides.

[kunal-mansukhani]

Follow-up: IDs 98, 102, and 104 were just merged from other PRs — please avoid those. Prefer 95, 97, 99–101, 103, or 120+.